diff --git a/ggml-quants.c b/ggml-quants.c index 55a9496d1..4a522e014 100644 --- a/ggml-quants.c +++ b/ggml-quants.c @@ -2340,6 +2340,138 @@ size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * return (n/QK_K*sizeof(block_q6_K)); } +// ====================== "True" 2-bit (de)-quantization + +void quantize_row_iq2_xxs_reference(const float * restrict x, block_iq2_xxs * restrict y, int k) { + (void)x; + (void)y; + (void)k; + assert(k % QK_K == 0); + //fprintf(stderr, "=========================== %s: not implemented\n", __func__); +} + +static const uint64_t iq2xxs_grid[256] = { + 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, + 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808, + 0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819, + 0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819, + 0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b, + 0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808, + 0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08, + 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b, + 0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819, + 0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08, + 0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808, + 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08, + 0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808, + 0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808, + 0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919, + 0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819, + 0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08, + 0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908, + 0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819, + 0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808, + 0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808, + 0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908, + 0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808, + 0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08, + 0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819, + 0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819, + 0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819, + 0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908, + 0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19, + 0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819, + 0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b, + 0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808, + 0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908, + 0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08, + 0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08, + 0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908, + 0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819, + 0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808, + 0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808, + 0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19, + 0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819, + 0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, + 0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b, + 0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08, + 0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808, + 0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908, + 0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b, + 0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819, + 0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08, + 0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08, + 0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808, + 0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b, + 0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b, + 0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908, + 0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819, + 0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808, + 0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908, + 0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b, + 0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808, + 0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b, + 0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b, + 0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808, + 0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19, + 0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908, +}; + +static const uint8_t ksigns_iq2xs[128] = { + 0, 129, 130, 3, 132, 5, 6, 135, 136, 9, 10, 139, 12, 141, 142, 15, + 144, 17, 18, 147, 20, 149, 150, 23, 24, 153, 154, 27, 156, 29, 30, 159, + 160, 33, 34, 163, 36, 165, 166, 39, 40, 169, 170, 43, 172, 45, 46, 175, + 48, 177, 178, 51, 180, 53, 54, 183, 184, 57, 58, 187, 60, 189, 190, 63, + 192, 65, 66, 195, 68, 197, 198, 71, 72, 201, 202, 75, 204, 77, 78, 207, + 80, 209, 210, 83, 212, 85, 86, 215, 216, 89, 90, 219, 92, 221, 222, 95, + 96, 225, 226, 99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111, + 240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255, +}; +static const uint8_t kmask_iq2xs[8] = {1, 2, 4, 8, 16, 32, 64, 128}; + +void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y, int k) { + assert(k % QK_K == 0); + const int nb = k / QK_K; + + uint32_t aux32[2]; + const uint8_t * aux8 = (const uint8_t *)aux32; + + for (int i = 0; i < nb; i++) { + + const float d = GGML_FP16_TO_FP32(x[i].d); + + for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { + memcpy(aux32, x[i].qs + 4*ib32, 2*sizeof(uint32_t)); + const float db = d * (0.5f + (aux32[1] >> 28)); + for (int l = 0; l < 4; ++l) { + const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]); + const uint8_t signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127]; + for (int j = 0; j < 8; ++j) { + y[j] = db * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); + } + y += 8; + } + } + } +} + +void quantize_row_iq2_xxs(const float * restrict x, void * restrict vy, int k) { + assert(k % QK_K == 0); + block_iq2_xxs * restrict y = vy; + quantize_row_iq2_xxs_reference(x, y, k); +} + +size_t ggml_quantize_iq2_xxs(const float * src, void * dst, int n, int k, int64_t * hist) { + assert(k % QK_K == 0); + (void)hist; // TODO: collect histograms + + for (int j = 0; j < n; j += k) { + block_iq2_xxs * restrict y = (block_iq2_xxs *)dst + j/QK_K; + quantize_row_iq2_xxs_reference(src + j, y, k); + } + return (n/QK_K*sizeof(block_iq2_xxs)); +} + //===================================== Q8_K ============================================== void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k) { @@ -7065,3 +7197,40 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri } #endif + +void ggml_vec_dot_iq2_xxs_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + assert(n % QK_K == 0); + + const block_iq2_xxs * restrict x = vx; + const block_q8_K * restrict y = vy; + + const int nb = n / QK_K; + + uint32_t aux32[2]; + const uint8_t * aux8 = (const uint8_t *)aux32; + + float sumf = 0.f; + for (int i = 0; i < nb; ++i) { + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const uint16_t * restrict q2 = x[i].qs; + const int8_t * restrict q8 = y[i].qs; + int32_t bsum = 0; + for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { + memcpy(aux32, q2, 2*sizeof(uint32_t)); + q2 += 4; + const uint32_t ls = 2*(aux32[1] >> 28) + 1; + int32_t sumi = 0; + for (int l = 0; l < 4; ++l) { + const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]); + const uint8_t signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127]; + for (int j = 0; j < 8; ++j) { + sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); + } + q8 += 8; + } + bsum += sumi * ls; + } + sumf += d * bsum; + } + *s = 0.125f * sumf; +} diff --git a/ggml-quants.h b/ggml-quants.h index 62c1df6cb..8dd911d41 100644 --- a/ggml-quants.h +++ b/ggml-quants.h @@ -165,6 +165,14 @@ typedef struct { } block_q8_K; static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding"); +// (Almost) "true" 2-bit quantization. +// Due to the need to use blocks as per ggml dsign, it ends up using +// 2.0625 bpw because of the 16-bit scale for each block of 256. +typedef struct { + ggml_fp16_t d; + uint16_t qs[QK_K/8]; +} block_iq2_xxs; +static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding"); // Quantization void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k); @@ -180,6 +188,7 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k); void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k); void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k); +void quantize_row_iq2_xxs_reference(const float * restrict x, block_iq2_xxs * restrict y, int k); void quantize_row_q4_0(const float * restrict x, void * restrict y, int k); void quantize_row_q4_1(const float * restrict x, void * restrict y, int k); @@ -194,6 +203,7 @@ void quantize_row_q4_K(const float * restrict x, void * restrict y, int k); void quantize_row_q5_K(const float * restrict x, void * restrict y, int k); void quantize_row_q6_K(const float * restrict x, void * restrict y, int k); void quantize_row_q8_K(const float * restrict x, void * restrict y, int k); +void quantize_row_iq2_xxs(const float * restrict x, void * restrict y, int k); // Dequantization void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k); @@ -209,6 +219,7 @@ void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k); void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k); void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k); +void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y, int k); // Dot product void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy); @@ -222,3 +233,4 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy); void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy); void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy); +void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy); diff --git a/ggml.c b/ggml.c index 62f0f18ef..adb387100 100644 --- a/ggml.c +++ b/ggml.c @@ -573,6 +573,17 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .vec_dot = ggml_vec_dot_q6_K_q8_K, .vec_dot_type = GGML_TYPE_Q8_K, }, + [GGML_TYPE_IQ2_XXS] = { + .type_name = "iq2_xxs", + .blck_size = QK_K, + .type_size = sizeof(block_iq2_xxs), + .is_quantized = true, + .to_float = (ggml_to_float_t) dequantize_row_iq2_xxs, + .from_float = quantize_row_iq2_xxs, + .from_float_reference = (ggml_from_float_t) quantize_row_iq2_xxs_reference, + .vec_dot = ggml_vec_dot_iq2_xxs_q8_K, + .vec_dot_type = GGML_TYPE_Q8_K, + }, [GGML_TYPE_Q8_K] = { .type_name = "q8_K", .blck_size = QK_K, @@ -2111,6 +2122,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { case GGML_FTYPE_MOSTLY_Q4_K: wtype = GGML_TYPE_Q4_K; break; case GGML_FTYPE_MOSTLY_Q5_K: wtype = GGML_TYPE_Q5_K; break; case GGML_FTYPE_MOSTLY_Q6_K: wtype = GGML_TYPE_Q6_K; break; + case GGML_FTYPE_MOSTLY_IQ2_XXS: wtype = GGML_TYPE_IQ2_XXS; break; case GGML_FTYPE_UNKNOWN: wtype = GGML_TYPE_COUNT; break; case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break; } @@ -7436,6 +7448,7 @@ static void ggml_compute_forward_add( case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ2_XXS: { ggml_compute_forward_add_q_f32(params, src0, src1, dst); } break; @@ -7700,6 +7713,7 @@ static void ggml_compute_forward_add1( case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ2_XXS: { ggml_compute_forward_add1_q_f32(params, src0, src1, dst); } break; @@ -7814,6 +7828,7 @@ static void ggml_compute_forward_acc( case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ2_XXS: default: { GGML_ASSERT(false); @@ -10455,6 +10470,7 @@ static void ggml_compute_forward_out_prod( case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ2_XXS: { ggml_compute_forward_out_prod_q_f32(params, src0, src1, dst); } break; @@ -10629,6 +10645,7 @@ static void ggml_compute_forward_set( case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ2_XXS: default: { GGML_ASSERT(false); @@ -10823,6 +10840,7 @@ static void ggml_compute_forward_get_rows( case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ2_XXS: { ggml_compute_forward_get_rows_q(params, src0, src1, dst); } break; @@ -11459,6 +11477,7 @@ static void ggml_compute_forward_alibi( case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ2_XXS: case GGML_TYPE_Q8_K: case GGML_TYPE_I8: case GGML_TYPE_I16: @@ -11533,6 +11552,7 @@ static void ggml_compute_forward_clamp( case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ2_XXS: case GGML_TYPE_Q8_K: case GGML_TYPE_I8: case GGML_TYPE_I16: @@ -18648,6 +18668,12 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i block_q6_K * block = (block_q6_K*)dst + start / QK_K; result = ggml_quantize_q6_K(src + start, block, n, n, hist); } break; + case GGML_TYPE_IQ2_XXS: + { + GGML_ASSERT(start % QK_K == 0); + block_iq2_xxs * block = (block_iq2_xxs*)dst + start / QK_K; + result = ggml_quantize_iq2_xxs(src + start, block, n, n, hist); + } break; case GGML_TYPE_F16: { int elemsize = sizeof(ggml_fp16_t); diff --git a/ggml.h b/ggml.h index 64f4e45e8..c55e598b4 100644 --- a/ggml.h +++ b/ggml.h @@ -339,6 +339,7 @@ extern "C" { GGML_TYPE_Q5_K = 13, GGML_TYPE_Q6_K = 14, GGML_TYPE_Q8_K = 15, + GGML_TYPE_IQ2_XXS = 16, GGML_TYPE_I8, GGML_TYPE_I16, GGML_TYPE_I32, @@ -373,6 +374,7 @@ extern "C" { GGML_FTYPE_MOSTLY_Q4_K = 12, // except 1d tensors GGML_FTYPE_MOSTLY_Q5_K = 13, // except 1d tensors GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors + GGML_FTYPE_MOSTLY_IQ2_XXS = 15, // except 1d tensors }; // available tensor operations: @@ -2067,6 +2069,7 @@ extern "C" { GGML_API size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist); GGML_API size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist); GGML_API size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist); + GGML_API size_t ggml_quantize_iq2_xxs(const float * src, void * dst, int n, int k, int64_t * hist); GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist);